CN114625886A - Entity query method and system based on knowledge graph small sample relation learning model - Google Patents

Abstract

The invention belongs to the technical field of knowledge graph data processing, and provides an entity query method and system based on a knowledge graph small sample relation learning model. The method includes acquiring the information interaction data to be queried and the relation-entity pairs contained therein, and encoding to obtain a vector to be queried; and performing feature encoding on the head and tail entity pairs contained in the vector to be queried based on the knowledge graph to obtain a corresponding triplet Representation; match the triple representation of the information interaction data to be queried with the triple representation of the pre-clustered groups of information cross-reference small samples to perform attention mechanism matching, and obtain the most similar group of information cross-reference small samples as the query. result.

Description

Entity query method and system based on knowledge graph small sample relation learning model

technical field

The invention belongs to the technical field of knowledge graph data processing, and in particular relates to an entity query method and system based on a knowledge graph small sample relation learning model.

Background technique

The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

At present, in unstructured environments such as homes and hospitals, service robots cannot perform efficient navigation and safely perform various service tasks because server humans do not really have the capabilities of autonomous learning, knowledge sharing, and emotional interaction like humans. Knowledge graphs (KGs) are the focus of research in the human-robot interaction problem of service robots. To further expand the coverage of KGs, traditional KGs completion methods require a large number of training instances (i.e. head-tail entity pairs) for each relation. Long-tailed relations are actually more common in KGs, and these newly added relations usually do not have many known training triples.

In practical applications, there are few human-computer interaction samples of the server person, and compared with the traditional knowledge graph learning method, the small sample knowledge representation learning needs to consider not only the difference in the number of reference samples, but also the need to consider Refer to the use of semantic information and result information between samples. None of the current algorithms consider the dynamic properties of learning entity triples, i.e., entities may exhibit different roles in task relations, while reference samples may contribute differently to query samples. Entity embedding methods, such as existing design modules to enhance entities and their local graph neighbors, do not take full advantage of supervised information. In the current technical solution, this problem is solved by learning the static representation of entities and references. The relationship is simply represented by the entity, but the feature relationship between the relationships is ignored, that is, the features between the relationships in the reference set may make different queries to the query. contribute.

In the process of research and development, the inventor found that the existing small-sample knowledge representation learning methods have huge shortcomings such as insufficient use of reference sample information and serious noise interference, and previous studies have not considered the dynamic attributes of samples combined with the fine-grained semantics of actual samples. , therefore, this reduces the accuracy of the actual server person's human-computer interaction question answering, making the server person's experience poor.

SUMMARY OF THE INVENTION

In order to solve the technical problems existing in the above-mentioned background art, the present invention provides an entity query method and system based on a knowledge graph small sample relation learning model, which can improve the accuracy of the human-computer interaction question and answer of the actual server person, so that the server person's Experience is better.

In order to achieve the above object, the present invention adopts the following technical solutions:

A first aspect of the present invention provides an entity query method based on a knowledge graph small sample relation learning model, which includes:

Obtain the information interaction data to be queried and the relation-entity pairs it contains, and encode to obtain the vector to be queried;

Based on the knowledge graph, the head and tail entity pairs contained in the vector to be queried are feature-encoded to obtain the corresponding triple representation;

Match the triple representation of the information interaction data to be queried with the triple representation of the pre-clustered groups of information cross-reference small samples to perform attention mechanism matching, and obtain the most similar group of information cross-reference small samples as the query result.

A second aspect of the present invention provides an entity query system based on a knowledge graph small sample relation learning model, which includes:

a query vector encoding module, which is used to obtain the information interaction data to be queried and the relation-entity pairs contained therein, and encodes to obtain the to-be-queried vector;

The triplet representation module is used to encode the head and tail entity pairs contained in the vector to be queried based on the knowledge graph to obtain the corresponding triplet representation;

The matching search module is used to match the triple representation of the information interaction data to be queried with the triple representation of the pre-clustered groups of information cross-reference small samples to perform attention mechanism matching to obtain the most similar group of information cross-references Small sample and as query result.

A third aspect of the present invention provides a computer-readable storage medium on which a computer program is stored, and when the program is executed by a processor, implements the above-mentioned entity query method based on a knowledge graph small sample relation learning model step.

A fourth aspect of the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, when the processor executes the program, the above-mentioned based Steps in an entity query method for a knowledge graph few-shot relational learning model.

Compared with the prior art, the beneficial effects of the present invention are:

The invention effectively utilizes the unique advantages of the attention mechanism and the entity coupling method, allocates the contribution weight in a finer-grained manner, optimizes the proportion of score elements, makes full use of the reference sample information, strengthens the entity embedding, and improves the human-computer interaction question-and-answer of the actual server person. The accuracy of the server makes the experience of the server better.

Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will become apparent from the description which follows, or may be learned by practice of the invention.

Description of drawings

The accompanying drawings forming a part of the present invention are used to provide further understanding of the present invention, and the exemplary embodiments of the present invention and their descriptions are used to explain the present invention, and do not constitute an improper limitation of the present invention.

Fig. 1 is an entity query method process framework based on a knowledge graph small sample relation learning model according to an embodiment of the present invention;

2 is a schematic diagram of a cloud service platform resource scheduling consumption model according to an embodiment of the present invention;

3 is a schematic diagram of the structure of an LSTM in a loop processor according to an embodiment of the present invention;

Fig. 4 is the NELL data set relational frequency visualization simulation effect diagram of the embodiment of the present invention;

Fig. 5 (a) is the characteristic clustering simulation result schematic diagram of the NELL data set relation r of the embodiment of the present invention;

5(b) is a schematic diagram of a feature clustering simulation result of the WiKi dataset relationship r according to an embodiment of the present invention;

6 is a schematic structural diagram of a feature cluster coding module according to an embodiment of the present invention;

7 is a schematic structural diagram of a transformer feature encoding module based on head and tail entity pairs according to an embodiment of the present invention;

8 is a schematic diagram of a specific structure of a matching scoring network module according to an embodiment of the present invention;

Fig. 9 (a) is the MRR experiment result analysis diagram of exploring the influence of sample quantity K on the predicted result according to the embodiment of the present invention;

Fig. 9 (b) is the Hits@10 experimental result analysis diagram of the influence of the sample number K on the predicted result according to the embodiment of the present invention;

Fig. 9 (c) is the Hits@5 experimental result analysis diagram of the influence of the exploration sample number K on the predicted result according to the embodiment of the present invention;

Fig. 9 (d) is the Hits@1 experimental result analysis diagram of the influence of the sample number K on the predicted result according to the embodiment of the present invention;

Fig. 10 (a) is the MRR experiment result analysis diagram of exploring the influence of the number of reference sample neighbors on the predicted result according to the embodiment of the present invention;

Figure 10(b) is an analysis diagram of the Hits@10 experimental result of exploring the influence of the number of reference sample neighbors on the prediction result according to an embodiment of the present invention;

Fig. 10 (c) is the Hits@5 experiment result analysis diagram of exploring the influence of the number of reference sample neighbors on the prediction result according to the embodiment of the present invention;

FIG. 10(d) is an analysis diagram of the Hits@1 experimental result of exploring the influence of the number of reference sample neighbors on the prediction result according to an embodiment of the present invention.

Detailed ways

The present invention will be further described below with reference to the accompanying drawings and embodiments.

It should be noted that the following detailed description is exemplary and intended to provide further explanation of the invention. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

It should be noted that the terminology used herein is for the purpose of describing specific embodiments only, and is not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly dictates otherwise, the singular is intended to include the plural as well, furthermore, it is to be understood that when the terms "comprising" and/or "including" are used in this specification, it indicates that There are features, steps, operations, devices, components and/or combinations thereof.

Example 1

Referring to FIG. 1, the present embodiment provides an entity query method based on a knowledge graph small sample relation learning model, which includes:

S101: Acquire the information interaction data to be queried and the relation-entity pairs contained therein, and encode to obtain a vector to be queried.

As shown in Figure 2, in the knowledge graph, an entity may be both the head entity of a triplet and the tail entity of multiple triples, that is, the entity may show different roles in the task relationship. Depending on the triple relationship, entities have different meanings in them. Therefore, when predicting the hidden relationship, the model can use its neighbor nodes, such as the relationship between the 5/6/7/8 nodes in Figure 4, to learn the relationship between the relationships.

For example: the triplet to be queried (Ming Yao, Work For, Chinese men's basketball team), where (Ming Yao, Chinese men's basketball team) is the head and tail entity pair, Yao Ming and Chinese men's basketball are the head and tail entities of the entity pair, respectively, Work For is the query relation vector to be clustered and encoded.

On the other hand, reference sets with different semantics may contribute differently to the query. For example, for the relationship "work for" in the query, the relationships "work for" and "work as" in the reference sample are more similar to the query sample than "subject to", "famousfor", etc., and the degree of influence when matching scores is affected. It will also be larger, and the model uses a fine-grained attention mechanism to highlight the contribution of strongly correlated samples.

In the specific implementation process, each group of information cross-reference small samples adopts unsupervised clustering, generates a preset number of center points, and performs clustering on the vector representation of each relationship in the reference small samples.

For example, through the K-means central clustering algorithm and cosine similarity calculation, N relational clusters in the reference sample are obtained, and each relational classification generates a public relation feature, so that the embedding data can be learned and weighted more finely.

In some specific implementation processes, a feature clustering coding module is used to cluster each group of information cross-reference small samples, as shown in FIG. 6 , for example. First, relational clustering is performed for the given reference sample data, and n known fact groups [r _i ] are obtained. The reference relationship-entity pairs in each grouping are coded to explore their degree of correlation. The module uses the following function f that satisfies the above properties:

是关系实体对(r_i,t_i)的特征表示，σ为激活函数，设置为σ＝tanh。in

is the feature representation of the relation entity pair (r _i , t _i ), σ is the activation function, set as σ=tanh.

A feedforward layer is then applied to encode the interactions in this tuple:

表示连接。where ω _c and b _c are learnable parameters,

Indicates connection.

After obtaining the relationship function f, when a query sample is input, the module encodes the relationship between the reference sample and the query sample and the neighbor head and tail entities respectively, and obtains:

where s and q are the vector representations of the relational neighbors between the reference sample and the query sample, respectively, where ^rs _i , h ^s _i , and t ^s _i are the feature representations of the relation and the neighbors of the head and tail entities in the reference sample, respectively, r ^qi _, h ^qi and ^{t qi} _are the feature representations of the relationship in the query sample and the neighbors of the head and tail _entities , respectively.

Through the correlation analysis between the query relationship and each relationship group [r _i ] _, when a query sample qi is input, the cluster coding module will perform cosine similarity analysis between the query sample and each relationship category, and analyze the relationship with the highest similarity. The category is given the highest weight. According to the relational neighbor vector representation of the reference sample and the query sample, the weight is assigned by the cosine similarity function:

The greater the correlation between the relationship of the query sample and a certain group, the higher the weight α _i . In this way, through the preliminary classification weighting, the query sample is classified into a certain reference sample relationship group with the greatest correlation.

Subsequently, the module performs fine-grained weight matching on query samples according to the correlation between task relations and reference relations. First define a metric function ψ, and calculate their correlation interaction representation by bilinear dot product:

ψ( _rs ,r _q )= _rs ^T Wr _q +b

Among them, r _s and r _q are the vector representations of the relationship between reference and query samples, respectively, and W and b are learnable parameters.

Under the initial weight of the weight α _i , the model further weights the similarity between the single sample in each group and the query sample to obtain the secondary weight β _ij as:

where m _n represents the nth reference sample in the mth relation grouping.

From this, the initial weight and secondary weight between the query sample and each reference sample can be obtained, and the attention parameters of the feature clustering encoding model based on the relation r can be obtained through the dot product operation:

The neighbor vector of the relationship is then weighted by attention, and the output of the cluster encoding model is obtained, that is, the pre-trained entity embedding, taking the head entity as an example:

where σ is the activation function. Set σ=tanh. The entity representation obtained by the feature clustering coding model based on the relation r in this way retains the individual attributes generated by the current embedding model, and according to the different roles played by different reference samples in the query sample, more fine-grained analysis of the reference sample and query The correlation between samples is described. The above formula also applies to the candidate tail entity t.

S102: Encode the head and tail entity pairs included in the vector to be queried based on the knowledge graph to obtain a corresponding triple representation.

For the reference samples in each relation grouping [r _i ], hide its relation grouping [r _i ], take each entity pair (h _ij , [r _i ], t _ij ) with task relation as a sequence X, and Coupling the entity embeddings with their role-aware neighbor embeddings, after building all input representations, feed them into the Transformer block stack, encode X and obtain the encoded triplet representation, sort the desired representation, and get the matching triplet Group.

Specifically, the feature encoding of the transformer feature encoding module based on the head-tail entity pair is adopted, and its operating mechanism is as follows:

The specific structure of the Transformer feature encoding module is shown in Figure 7. Module Given a triple in task r, i.e. (h _ij , [r _i ], t _ij ) ∈ Dr, for each reference sample in relation grouping [r _i ], we hide its relation grouping [r _i ] ], take each entity pair (hi _ij , [ri ], t _ij ) with task relation as a sequence X ₌ (x1, x2, x3), where the first/last element is the head/tail entity, In the middle is the hidden task relationship. Through the clustering module, we represent the reference sample head and tail entities as feature embeddings after training. To enhance the entity embedding, taking the head entity h as an example, the module simultaneously couples the entity embedding h with its role-aware neighbor embeddings. h can be expressed as:

F _C (h)=σ[ω ₁ h ^s _i +ω ₂ F _& (h ^s _i )]

where σ is the activation function. Set σ=relu, and W1 and W2 are learnable parameters.

After building all input representations, the encoding module feeds feature embeddings into the stack of Transformer blocks, encodes X and obtains:

Z ^k _i =Transformer(Z ^k-1 _i ),k=1,2,...,L

where Z ^k _i is the hidden state of the entity pair after the kth layer, and L is the number of hidden layers of the Transformer.

The final hidden state Z ^L ₂ serves as the entity's expected representation of the relation, as the relation embedding of the new triple when matching scores. This representation encodes the semantic role of each entity, thereby helping to identify the fine-grained meaning of task relationships associated with different pairs of entities.

In order to use the limited reference samples in a finer-grained manner and make the reference samples more relevant to the query play a greater role, the module samples and sorts the hidden state Z ^L ₁ /Z ^L ₃ of the entities finally output by the Transformer, according to the weight of each relationship grouping Different, according to the Euclidean space distance formula to sort the sample correlation:

g ^h _i =Z ^L ₁ (S _i ), g ^t _i =Z ^L ₃ (S _i )

Sort1={MIN[F(h ^q _i ,t ^q _i )],(h ^q _i ,r ^q _i ,t ^q _i )∈G',r ^q _i ∈[r _i ]}

Where g ^h _i /g ^t _i is the hidden state Z ^L ₁ /Z ^L ₃ of the entity pair that the reference sample enters into the final output of the Transformer. The smaller the distance between the reference sample entity pair and a certain grouped query sample entity pair, the greater the correlation between the query sample and the group of reference samples, and the greater the probability of being in the same class when clustering. Sort1 outputs the minimum distance between the reference sample entity pair and the query sample entity pair, and takes the first n groups of output sort corresponding groups as the reference embedding for the next step.

S103: Perform attention mechanism matching between the triple representation of the information interaction data to be queried and the triple representation of the pre-clustered groups of information cross-reference small samples to obtain the most similar group of information cross-reference small samples and use them as a query result.

In a specific implementation process, the similarity is represented by Euclidean distance. In the process of matching the attention mechanism, the cosine similarity function is used to assign weights between the triple representation of the information interaction data to be queried and the triple representation of each group of information interaction reference samples.

In the process of obtaining the most similar group of information cross-reference small samples, according to the different weights occupied by each relationship group, the correlation between the information interaction data to be queried and each group of information cross-reference small samples is sorted according to the Euclidean space distance formula.

Multi-step matching is performed using the LSTM in the loop processor in Figure 3. The input of the LSTM hidden layer includes the state c _t-1 of the hidden layer at the previous moment, the output vector h _t-1 of the hidden layer at the previous moment and the sequence input x _t at the current moment. The forget gate of LSTM controls the memory of the state of the previous storage unit, and determines how much information in the state c _t-1 of the storage unit at the previous moment can be transferred to the current moment c _t ; the input gate determines how much information in the current sequence input x _t It can be saved to the current time c _t ; the output gate obtains the current time output h _{t based on the new state c t .} The update method of LSTM can be expressed as:

f _t =σ(W _xf x _t +W _hf h _t-1 +b _f )

i _t =σ(W _xi x _t +W _hi h _t-1 +b _i )

o _t =σ(W _xo x _t +W _ho h _t-1 +b _o )

h _t =o _t ·tanh(c _t )

为当前时刻积累的状态信息，W表示不同门所对应的权重系数矩阵，b表示的偏置项，σ和tanh分别代表sigmoid激活函数和tanh激活函数。In the formula, c _t is the state information of the storage unit at the current moment,

is the state information accumulated at the current moment, W represents the weight coefficient matrix corresponding to different gates, b represents the bias term, σ and tanh represent the sigmoid activation function and the tanh activation function, respectively.

In the specific implementation, the matching scoring network module is used to perform matching query, and its operating mechanism is as follows:

The specific structure of the matching score network module is shown in Figure 8. Embeddings of the reconstructed new relational query samples and the most similar grouped reference samples are obtained. The matching score network module matches each query triplet ( _hi ,r _new ,t _i ) with the reference set (h,[r _i ],t) to predict the score result.

To measure the similarity between two vectors, we use a recurrent processor F(m) to perform multi-step matching. The t-th processing step is expressed as follows:

μ ₁ =( _hi ,r _new ,t _i ),μ ₂ =(h,[r _i ],t)

g _t =g't+μ ₁

where RNN _match is an LSTM cell with input μ ₁ , hidden state g _t and cell state c _t . The last hidden state g _T after the T "processing" step is the refined embedding of the query triplet: μ ₁ =g _T . _The matching score module uses the inner product between _μ1 and μ2 as the similarity score for the subsequent ranking optimization process.

This embodiment is simulated and verified.

During the training process of the model, a large amount of raw data will be generated. There are a lot of missing and noise in these raw data, which seriously affects the quality of the data and causes certain troubles in mining effective information. Applying some methods, such as data cutting, can improve the data. the quality of. Data preprocessing divides the training data in the original public data set NELL data set. According to the frequency of occurrence, the relationship in the training set is divided into three partitions: low, medium and high, and the corresponding orders of magnitude are [0,200), [200,400), [400,400 respectively. +]. Visual analysis is performed according to the frequency of relationship occurrence. The results are shown in Figure 4. The relationship frequency in the data set conforms to the long-tail distribution, which meets the requirements of model training.

Figure 5(a)-Figure 5(b) shows the simulation results of the feature clustering coding grouping of the reference sample triplet relationship r. We perform K-means center clustering on the public dataset NELL and WiKi dataset respectively. And carry out visual analysis to obtain the optimal clustering results when the number of center points K is different. Taking K=6 and K=7 as an example, it can be observed that the number of reference samples in the NELL and WiKi datasets is N=3 and N=5, respectively, and the clustering degree is relatively high.

Experiments are performed on the public datasets NELL dataset and WiKi dataset, selecting those relations that do not have too many triples as one-time task relations, picking the latest dump relations, and removing these inverse relations. The dataset selected relations with less than 500 but more than 50 triplets as one-off tasks.

The FANC model is trained with multiple baseline models under the same parameter pre-trained model, and for all implemented few-shot learning methods, we initialize entity embeddings via TransE. Random sampling and fixation of entity neighbors before model training. The models are trained using the relationships and their entity pairs in the training data of the public datasets NELL and WiKi, and the models are adjusted and evaluated using the relationships in the validation data and test data, respectively. The experiments use top-k hit ratio (Hitsk) and mean reciprocal rank (MRR) to evaluate the performance of different methods. k is set to 1, 5 and 10.

Taking the NELL dataset as an example, the performance of all models on it is as follows:

The experimental results clearly prove that compared with the traditional KG embedding method, the FANC model achieves better performance on both datasets, and the small-sample relation learning network model based on the attention clustering knowledge graph is more suitable for solving small-scale problems. Sample question.

Taking the NELL dataset as an example, the experimental results to explore the influence of the sample size K on the prediction results are shown in Figure 9(a)-Figure 9(d). Under different K, the FANC model outperforms all baseline models, showing the effectiveness of the new model in small-sample scenarios. When the sample K increases to a certain extent, the effect of the traditional model stagnates, and the FANC model still shows an increase in effect, indicating that in the scene of small samples, a larger reference set does not always achieve better performance. Noise is introduced because small sample scenarios make performance sensitive to available references. However, the transformer coding module in the new model selects the n groups of samples with the highest correlation, which significantly reduces the interference of irrelevant reference data and the introduction of noise, and the model prediction results are optimized and improved.

Taking the NELL dataset as an example, the experimental results to explore the influence of the number of reference sample neighbors on the prediction results are shown in Figure 10(a)-Figure 10(d). For each entity, traditional models that encode more neighbors sometimes yield worse performance. The reason is that for some entity pairs, there are some local connections that are irrelevant and provide noisy information to the model. For the FANC model, triples are modeled by a clustered neighbor encoder, which is clustered by relation to identify their task-oriented roles, and given corresponding weights according to their contributions. Both the reference set and the query set can capture the fine-grained semantic meaning of neighbors. Neighbor vectors with high correlation have more weight, so as to present a more expressive representation and effectively avoid noise interference.

The small sample relationship learning network process framework based on the attention clustering knowledge graph proposed in this embodiment utilizes the relationship clustering algorithm and the coupled sorting algorithm to optimize feature triples in a finer-grained manner, which improves the accuracy of the model prediction results and the algorithm convergence speed. The high efficiency and low cost of the algorithm are better verified by simulation experiments.

The advantages of the present invention are:

(1) The present invention proposes a new small sample knowledge graph representation learning process framework. The core of the similarity metric learning between query and reference samples is two encoding functions: F(&) and F(T). Therefore, for any query relation r _q , as long as there is a known fact group [r _i ], through the entity and relation encoding module, into the matching scoring network, the model can predict the test triplet (h _q ,r _q ,t _q ) ) and (h,[r _i ],t) matching scores to optimize the accuracy of prediction results.

(2) The present invention proposes a new feature clustering coding model based on relation r, which uses the entity representation obtained by the clustering model to retain the individual attributes generated by the current embedding model, and does not play the role of different reference samples in the query sample. The role of the sample is different, and the correlation between the reference sample and the query sample is described in a finer-grained manner.

(3) The present invention innovatively constructs an expectation sorting module in the transformer coding model, refers to the correlation between the sample entity pair and a certain grouped query sample entity pair, and selects the n groups of samples with the highest correlation degree, which significantly reduces the interference of irrelevant reference data. Noise is introduced, and the model prediction results are optimized and improved.

(4) The present invention effectively utilizes the unique advantages of the attention mechanism and the entity coupling method, assigns contribution weights in a finer-grained manner, optimizes the proportion of score elements, makes full use of reference sample information, and strengthens entity embedding.

Embodiment 2

This embodiment provides an entity query system based on a knowledge graph small sample relation learning model, which specifically includes the following modules:

a query vector encoding module, which is used to obtain the information interaction data to be queried and the relation-entity pairs contained therein, and encodes to obtain the to-be-queried vector;

The triplet representation module is used to encode the head and tail entity pairs contained in the vector to be queried based on the knowledge graph to obtain the corresponding triplet representation;

The matching search module is used to match the triple representation of the information interaction data to be queried with the triple representation of the pre-clustered groups of information cross-reference small samples to perform attention mechanism matching to obtain the most similar group of information cross-references Small sample and as query result.

It should be noted here that each module in this embodiment corresponds to each step in Embodiment 1 one by one, and the specific implementation process thereof is the same, which is not repeated here.

Embodiment 3

This embodiment provides a computer-readable storage medium on which a computer program is stored, and when the program is executed by a processor, implements the steps in the above-mentioned entity query method based on a knowledge graph small sample relation learning model.

Embodiment 4

This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the program, the above-mentioned knowledge graph-based small Steps in an entity query method for a sample relational learning model.

The present invention is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the invention. It will be understood that each flow and/or block in the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to the processor of a general purpose computer, special purpose computer, embedded processor or other programmable data processing device to produce a machine such that the instructions executed by the processor of the computer or other programmable data processing device produce Means for implementing the functions specified in a flow or flow of a flowchart and/or a block or blocks of a block diagram.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims (10)

1. An entity query method based on a knowledge graph small sample relation learning model is characterized by comprising the following steps:

acquiring information interaction data to be queried and a relation-entity pair contained in the information interaction data, and coding to obtain a vector to be queried;

performing feature coding on head and tail entity pairs contained in a vector to be queried based on a knowledge graph to obtain corresponding triple representations;

and carrying out attention mechanism matching on the triple representation of the information interaction data to be inquired and the triple representation of each group of information interaction reference small samples clustered in advance to obtain the information interaction reference small sample of the most similar group and using the information interaction reference small sample as an inquiry result.

2. The entity query method based on the knowledge-graph small-sample relation learning model as claimed in claim 1, wherein each group of information interaction reference small samples adopts unsupervised clustering to generate a preset number of central points, and the vector representations of each relation in the reference small samples are clustered.

3. The method for entity query based on knowledge-graph small-sample relation learning model according to claim 1, characterized in that the similarity is characterized by Euclidean distance.

4. The entity query method based on the knowledge-graph small-sample relation learning model as claimed in claim 1, wherein in the process of performing attention mechanism matching, weight distribution between the triple representation of the information interaction data to be queried and the triple representation of each group of information interaction reference small samples is performed through a cosine similarity function.

5. The entity query method based on the knowledge-graph small-sample relation learning model as claimed in claim 1, wherein in the process of obtaining the information cross-reference small samples of the most similar group, the information cross-reference data to be queried and the relevance of each group of information cross-reference small samples are sorted according to the Euclidean spatial distance formula according to the different weights occupied by each relation group.

6. An entity query system based on a knowledge graph small sample relation learning model is characterized by comprising:

the query vector encoding module is used for acquiring information interaction data to be queried and relation-entity pairs contained in the information interaction data, and encoding to obtain a vector to be queried;

the triple representation module is used for carrying out feature coding on head and tail entity pairs contained in the vector to be queried based on a knowledge graph to obtain corresponding triple representation;

and the matching search module is used for performing attention mechanism matching on the triple representation of the information interaction data to be inquired and the triple representation of each group of information interaction reference small samples clustered in advance to obtain the information interaction reference small sample of the most similar group and using the information interaction reference small sample as an inquiry result.

7. The system of claim 6, wherein in the matching and searching module, in the process of performing attention mechanism matching, weight distribution between the triple representation of the information interaction data to be queried and the triple representation of each set of information interaction reference small samples is performed through a cosine similarity function.

8. The system of claim 6, wherein in the matching and searching module, in the process of obtaining the information cross-reference small samples of the most similar group, the weights of the information cross-reference small samples of the most similar group are different according to the relation groups, and the relevance of the information cross-reference small samples and the information cross-reference data to be queried is sorted according to the Euclidean spatial distance formula.

9. A computer-readable storage medium, on which a computer program is stored, which program, when being executed by a processor, carries out the steps of the method for querying an entity based on a knowledge-graph small-sample relation learning model according to any one of claims 1 to 5.

10. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor when executing the program implements the steps in the method for entity query based on a knowledge-graph small-sample relationship learning model according to any one of claims 1-5.

Images